关键词: 流形学习, 局部线性嵌入, 特征融合, 汉语方言辨识

Abstract: A feature extraction method based on manifold learning is proposed. Focusing on the high dimension among the spectrogram feature spaces, the locally linear embedding algorithm is applied to dimension reduction, and the feature is fused with Mel Frequency Cepstrum Coefficient（MFCC）, and the fusion result is used as a new feature in Chinese dialect identification system. The manifold learning method is effectively applied to Chinese dialect identification system. Simulation results show that the Locally Linear Embedding（LLE） algorithm can obtain the intrinsic law of dialects, and the fused characteristics can effectively improve the correct recognition rate of Chinese dialect identification.

Key words: manifold learning, Locally Linear Embedding（LLE）, feature fusion, Chinese dialect identification